The goal of this proposal is to develop knowledge-based methods for interactively extracting (or segmenting), anatomic objects from medical images. The National Library of Medicine has recently released a Request for Proposals, the "Visible Human Project'", which calls for the creation of a digital image library of volumetric data representing a complete normal adult human male and female cadaver. Such an image library will find wide application in all areas of medicine. However, as the NLM long-range plan recognizes, the raw image data itself will be of limited use until the relevant anatomic structures have been segmented and classified. This "segmentation problem" is a major bottleneck to automated image analysis in all areas of medicine, and is becoming critical as the rapidly increasing number of medical images threatens to overwhelm human interpreters. The two fundamental hypotheses behind my approach to this problem are 1)spatial knowledge of the shape and range of variation of anatomic objects is essential for segmentation, and 2) segmentation is a difficult problem that is not likely to be solved in the near future, so any useful system must be interactive. In previous work I have developed a representation, called geometric constraint networks, that is able to capture not only the shape, but also the range of variation of many biological objects ranging from proteins to organs. I have implemented this representation in an interactive program for segmentation of organs on 2-D medical images, and a preliminary evaluation has shown that shape knowledge can speed up segmentation time by as much as a factor of ten. In the current proposal I plan to continue this work. The specific aims are 1) to develop improved methods for representing spatial knowledge of anatomy based on geometric constraint networks 2) to develop a series of interactive prototypes for segmentation of 2-D and 3-D image datasets, and 3) to evaluate the efficacy of these prototypes in two practical applications, reconstruction of anatomy from serial cadaver sections, and radiation treatment planning. Fulfillment of these specific aims will result not only in useful tools that reduce the current image segmentation bottleneck, but also in representations that form the basis for a spatial knowledge base of anatomy. Because of the fundamental roles that anatomy and structural biology play within the medical sciences, such a knowledge base will have wide applicability in clinical medicine, teaching and basic research.